Evaluation

Measuring what systems can do and how reliably, at every layer: benchmarks, red-teaming, production monitoring.

Cross-cutting concern, spans every layer · Draft, awaiting expert review, reviewed 2026-10-14

D1 · StoryFor young children and total beginners

Predict: does a perfect test score always mean the AI is good?

Evaluation means testing. Builders give the AI lots of questions where they already know the right answer, and count how many it gets right.

They test every part, not only the answers: how much power the machines use, how fast they work, and whether the AI really helps people.

A test can fool you. If the AI has seen the questions before, a high score proves very little.

Like: A school test: useful, but only if the questions are new to the student.

D2 · PictureFor curious beginners

Every layer has its own yardstick. Data centres report power usage effectiveness (PUE). Chips and clusters are compared on standard workloads that time how fast a system trains or answers. Models are scored on test sets of questions. Apps are judged by what they change for real people.

There are three kinds of judge: code (exact answers, passing unit tests), people (ratings and side-by-side votes), and another model acting as judge, which is cheap but has biases of its own.

One number hides a lot. Holistic evaluations score models on accuracy, calibration, robustness, fairness, bias, toxicity and efficiency side by side, because improving one can hurt another.

Agents need repeat trials. An agent that solves a task once may fail the same task next time, so reliability is measured as passing every one of several attempts (pass^k).

Like: A school test: useful, but only if the questions are new to the student.
Where it breaks: A student who passes usually keeps the skill. A model's score can change with small changes to the wording or order of questions, and a model may act differently when it can tell it is being tested.

D5 · FrontierFor experts

Test sets leak. Once questions are published they can end up in training data (contamination) and inflate scores; fresh, private or regularly refreshed tests reduce this.

Models increasingly notice when they are being tested. The 2026 International AI Safety Report warns that pre-deployment testing is becoming less reliable as a result.

Lab and field results disagree. In customer support, an AI assistant raised issues resolved per hour by 15% on average; in a randomised trial with experienced open-source developers, tasks took 19% longer. Task, setting and skill level matter.

Model judges are convenient but show position, verbosity and self-enhancement biases, so their scores need checking against people. Large public arenas collect human side-by-side votes instead, which reflect what voters prefer rather than what is correct.

On every layer

Layer 9 · Applications & Interfaces: The real test is what changes for people: in a 2025 randomised trial, experienced developers took 19% longer with AI tools, though they felt faster. (Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (METR))
Layer 8 · Agents: Agents are run on the same task several times: in one 2024 benchmark, leading agents solved fewer than half the tasks, and in one domain under a quarter on all eight tries. (τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains)
Layer 7 · Orchestration & Tooling: Retrieval pipelines are tested in parts: did the search find relevant passages, and did the answer stay faithful to them? (Ragas: Automated Evaluation of Retrieval Augmented Generation)
Layer 6 · Inference & Optimization: Faster serving must not change the answers: compression methods are judged by how little accuracy they lose against the uncompressed model. (GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers)
Layer 5 · Foundation Models: Models are scored on many test sets at once, and on more than accuracy: calibration, robustness, fairness, bias, toxicity and efficiency. (Holistic Evaluation of Language Models (HELM))
Layer 4 · Data: A dataset is judged by training the same small model on it and scoring the result, so different filtering choices can be compared fairly. (DataComp-LM: In search of the next generation of training sets for language models)
Layer 3 · Cloud & Infrastructure: Clusters are compared by the wall-clock time they take to train a set model to a target quality, with the number of accelerators listed. (MLPerf Training benchmark (MLCommons))
Layer 2 · Compute Hardware: Chips are compared on standard workloads that measure how fast a whole system answers requests, and can also record its power draw at the wall. (MLPerf Inference: Datacenter benchmark (MLCommons))
Layer 1 · Physical & Energy: Data centres measure their overhead with power usage effectiveness (PUE): total facility energy divided by the energy used by the IT equipment, where 1.0 would mean no overhead. (Best Practices Guide for Energy-Efficient Data Center Design (US DOE FEMP / NREL), PUE definition)

Examples as of 2026-10-14

MLPerf (MLCommons)
HELM
LMArena (formerly Chatbot Arena)
SWE-bench
OSWorld
τ-bench